A coding agent on your own API key means three things: you choose the model, you pay the provider's price with no markup, and your requests go to the provider you picked. This is how to set that up in VS Code — or Cursor, Kiro or Windsurf — with the Atlarix extension, start to finish.
1. Install the extension
In your editor's Extensions view, search for Atlarix and install it — it comes from the Visual Studio Marketplace in VS Code and from Open VSX in Cursor, Kiro and Windsurf. It brings its own copy of the agent, so there is no CLI or runtime to install separately.
2. Sign in once
Open the Atlarix view (the ant in the activity bar, or ⌘⌥A / Ctrl+Alt+A) and sign in. An Atlarix account is free; it is what keeps your chats, settings and keys the same across the editor, the desktop app and the CLI.
3. Connect your provider
Open the ⋯ menu at the top of the chat and choose Account, usage and billing. Under Your own keys, press Connect a provider…, pick the provider — OpenAI, Anthropic, Google, DeepSeek, OpenRouter, or another from the list — and paste the key. The prompt links to the page where that provider issues keys. The key is stored by Atlarix on your computer, and the Atlarix app and CLI on the same computer use it too.
4. Pick the model
The model picker sits just under the message box. Your key's models now appear in it alongside Atlarix Auto and the Atlarix Core models; choose one and it applies to this chat. Different chats can use different models — a fast cheap one for questions, a stronger one for a refactor.
5. Work as normal
Every mode works on your own key: Build, Explore, Plan, Debug and Review. Edits open in the editor's diff for your Yes or No, commands wait for approval, and plans tick off as they are built. The account panel shows what this chat has cost on your key so far.
What it costs, and where requests go
Atlarix adds nothing to your own key's usage: you pay the provider exactly what they charge. Requests on your key go from your computer to the provider, not through Atlarix. If the provider rejects the key — revoked, mistyped or out of credit — the chat says so, so you know to fix the key rather than retry.
Which model to start with
There is no single answer, and the honest advice is to try two. Frontier models from OpenAI, Anthropic and Google are the strongest at long multi-step changes; DeepSeek and GLM are far cheaper per token and very capable for most everyday work; OpenRouter gives you one key for many providers. Because the picker is per chat, comparing them on your own code costs one extra chat.
The same key works in Zed and JetBrains IDEs, which run Atlarix over the Agent Client Protocol — see the editors guide. And if you would rather not manage keys at all, Atlarix Auto in the same picker chooses a model for each step for you.
Originally published at atlarix.dev.
Top comments (0)